Patent-linked AI exposure in China nudges workers into short-range job switches, but the type of AI matters: labor-augmenting AI is associated with longer-distance occupational moves and comparatively better post-transition fit, whereas labor-saving AI predicts worse skill-match and longer hours for movers.

Main Finding

Instrumented occupational exposure to AI in China has asymmetric effects depending on whether AI is labor-saving (routine-task substituting) or labor-augmenting (non-routine-task complementary). Labor-augmenting exposure raises medium- and long-distance occupational switching (including moves across major occupational groups) and generates positive origin–destination sorting among switchers; labor-saving exposure mainly increases short-distance (within-group) moves and is associated with worse post-transition conditions (longer hours and lower skill-match). Less-educated workers face a barrier to long-distance transitions when originating in high labor-saving-exposed occupations, pointing to distributional risks.

Key Points

• Dataset and scope

  • China Labor Force Dynamics Survey (CLDS) 2012–2018; 2,446 worker-year observations with valid five-digit occupation codes.
  • Occupational task corpus: 9,334 task statements from the 2015 Chinese Standard Classification of Occupations.
  • Chinese AI patents: patyee database (2010–2018); U.S. AIPD used for instrument (machine learning, computer vision, NLP subfields).

• Measurement

  • Occupation-level AI exposure = semantic similarity between occupational task texts and patent texts.
  • Decomposed into routine-task (labor-saving) and non-routine-task (labor-augmenting) components.
  • Embeddings: paraphrase-multilingual-MiniLM-L12-v2 (384-d). Retrieval via FAISS (K=50); sparsification at 60th percentile. Task labeling (routine vs non-routine) done with GPT-3.5 and validated.

• Identification

  • 2SLS IV: interaction of 2010 baseline occupation exposure with lagged cumulative U.S. patent stocks (USPTO AIPD) in the three subfields (ML, CV, NLP). This leverages international patent supply shocks conditional on occupation-level baseline task-patent relatedness.
  • Sensitivity checks: alternative retrieval caps and sparsification thresholds (K ∈ {30,50,70}; τ ∈ {0.50,0.60,0.70,0.80}) and model variants — results robust to these choices.

• Core empirical results (2SLS)

  • Total AI exposure: +1 log unit → +3.6 percentage points (ppt) minor-group switching; +3.4 ppt intermediate-group switching; no significant major-group effect.
  • Decomposed effects:
    • Labor-augmenting exposure: significant positive gradient with distance — +4.6 ppt (intermediate), +6.9 ppt (major).
    • Labor-saving exposure: IV coefficients statistically indistinguishable from zero for full-sample switching incidence; effects concentrated at short distances in some subgroups.
  • Heterogeneity:
    • Short-distance switching increases concentrated among women.
    • Among less-educated workers, origin labor-saving exposure predicts LOWER major-group mobility while labor-augmenting predicts HIGHER — suggesting a labor-saving–related barrier to long-distance upgrading.
  • Sorting among switchers:
    • Labor-augmenting: positive but attenuated origin–destination sorting across all switch samples.
    • Labor-saving: positive sorting at minor-group level, ~0 at intermediate, negative at major-group level.
  • Post-transition job conditions (among switchers):
    • Higher origin labor-saving exposure → longer weekly hours (minor/intermediate) and lower skill-match satisfaction (all samples).
    • Positive associations for origin labor-augmenting exposure are sample-specific and not uniform.

• Interpretation & caveats

  • The exposure index measures potential applicability (frontier exposure) rather than realized workplace AI adoption.
  • Complementary-upgrading (labor-augmenting enabling access to more cognitive jobs) is posited as a candidate mechanism but not directly observed; worker selection and unobserved human-capital accumulation remain alternative explanations.
  • Instrument validity relies on relevance and exclusion (U.S. patent supply shocks affecting Chinese occupation-level exposure only via technology diffusion pathways captured by baseline relatedness).

Data & Methods

• Microdata: CLDS (2012–2018), ages 16–65, requiring valid five-digit occupational codes for current and previous job; final N = 2,446 worker-year observations.
• Occupational tasks: 9,334 Chinese task statements; classified routine vs non-routine using GPT-3.5 with human/cross-model validation.
• Patent corpora:
  • Chinese patents (patyee) for exposure construction (titles+abstracts, 2010–2018) with 93 AI-related keywords.
  • USPTO AIPD (machine learning, computer vision, NLP) used to form lagged cumulative patent-stock instrument.
• Exposure construction:
  • Encode tasks and patents with multilingual sentence-transformer (384-d), compute cosine similarities.
  • For each task, retrieve top-K patents (K=50) with FAISS; sparsify at 60th percentile; sum truncated similarities to form task-level scores; aggregate across routine and non-routine tasks to create occupation-level labor-saving and labor-augmenting scores; take ln(1 + ξ) for estimation.
• Outcome definitions:
  • Occupational switching measured at three nested distances: minor-group (any digit change), intermediate-group (first three digits), major-group (first digit).
  • Post-transition conditions among switchers: weekly working hours (job demands) and self-reported skill-match satisfaction (person–job fit).
• Empirical approach:
  • Main estimates: 2SLS models where exposure (total, labor-saving, labor-augmenting) is instrumented by baseline 2010 occupation exposure × lagged U.S. patent-stock growth; controls include individual and city covariates and propensity-score matching in supporting analyses.
  • Robustness: parameter grid for retrieval and sparsification; GLM variant; results stable.

Implications for AI Economics

• Importance of decomposing AI exposure: Aggregating exposure masks opposite effects of augmentation and substitution. Empirical and policy work should distinguish labor-augmenting and labor-saving channels.
• Mobility mechanisms and reallocation dynamics:
  • Labor-augmenting AI appears to facilitate longer-distance occupational upgrading and positive origin–destination sorting, consistent with complementarity expanding feasible cognitive destinations (or portability of non-routine skills).
  • Labor-saving AI induces defensive, short-distance moves and worsens post-transition job conditions, suggesting reallocation may be costly and quality-degrading for affected workers.
• Distributional concerns:
  • Less-educated workers in high labor-saving occupations face barriers to long-distance upgrading, implying rising mismatch and inequality risks; targeted reskilling and transition support should prioritize these groups.
• Measurement and identification practice:
  • Patent-text semantic-similarity methods are feasible in non-English contexts and can be combined with international patent flows as instruments to address endogeneity of frontier exposure.
  • Researchers should be explicit about the distinction between frontier exposure (potential applicability) and realized adoption; combining patent-based indices with firm-level adoption data would strengthen causal claims about realized effects.
• Modeling guidance for theory:
  • Models of AI and labor should embed two channels (substitution vs augmentation), allow channel-specific effects on switching costs and occupational distance, and account for sorting and post-transition job quality (hours, fit), not just employment counts.
• Policy levers:
  • Invest in reskilling that targets occupations with high labor-saving exposure and low education levels.
  • Provide organizational support and mobility assistance (career counseling, bridging training) to reduce the long-distance transition barrier.
  • Monitor post-transition outcomes (hours, satisfaction) as well as placement rates to evaluate the welfare implications of AI-driven reallocation.

Limitations to note for applied work: the study identifies associations using patent-based frontier exposure and an international-patent instrument; it cannot observe firm-level AI use or directly confirm the complementary-upgrading learning mechanism. Sample size is modest and China-specific, so replication in other contexts and with realized-adoption measures would be valuable.